From Residuals to Reasons: LLM-Guided Mechanism Inference from Tabular Data

Background

The article addresses the challenge of integrating prediction accuracy with explanatory power in machine learning models, particularly for scientific applications. Traditional statistical models are adept at structured data but lack interpretability. Current interpretability methods primarily identify important features without explaining interactions or refining explanations iteratively. Directly asking an LLM to predict targets forces it to explore a vast output space, which is impractical and ineffective.

Key Points

MARICL proposes a novel approach by anchoring predictions with a base model and having LLM agents analyze where the base-model fails. These agents hypothesize missing structures from high-residual examples provided in context and produce explicit correction terms refined through multi-turn textual gradient optimization. Specifically, MARICL was tested across nine diverse benchmarks: scientific, biomedical, socioeconomic, and synthetic data.

Iterative Textual Refinement

MARICL's process involves:

1. Base Model Prediction: The base model generates initial predictions.
2. Residual Analysis: LLM agents identify where the base-model fails by analyzing high-residual examples provided in context.
3. Hypothesis Formulation: Agents hypothesize missing structures that could explain discrepancies.
4. Correction Term Generation: Explicit correction terms are produced and refined through iterative textual optimization.

Consistent Improvement

Across all tested datasets, MARICL improves over its base model's performance consistently. This improvement is significant in demonstrating the effectiveness of integrating LLM agents to correct for errors made by base models.

Significance

The success of MARICL lies in its ability to produce mechanistic generalization. To test this, formulas learned on one experimental batch of the Cell-Free Protein dataset were frozen and applied to held-out batches. The results showed that within the same reagent protocol, these formulas improved predictions in over 92% of cases; however, they failed systematically when applied across a different protocol. This demonstrates that MARICL captures real structural information rather than mere batch-specific noise.

Key Insights:

• Mechanistic Generalization: The success boundary aligns with the biochemistry, indicating that the corrections reflect real structure and not just random fluctuations.
• Iterative Textual Optimization: This process allows for a more nuanced understanding of data, enhancing both prediction accuracy and interpretability.
• Applicability Across Domains: MARICL's approach is broadly applicable across various scientific fields, including biomedical, socioeconomic, and synthetic settings.

In summary, MARICL offers a robust framework to enhance the explanatory power of machine learning models in complex scientific applications by leveraging LLM agents for iterative correction. This method not only improves prediction accuracy but also demonstrates mechanistic generalization, providing deeper insights into the underlying data structures.